Pixels to pasture: Using machine learning and multispectral remote sensing to predict biomass and nutrient quality in tropical grasslands

The livestock sector in rural Colombia is critical for employment and food security but is heavily affected by climate and its change. There is a need for solutions to address key challenges arising from vulnerabilities that impact the productivity and sustainability of forages and the livestock sector. Increasing the yields of forage crops can improve the availability and affordability of livestock products while also easing the pressure on land resources. This study aims to develop remote sensing-based approaches for forage monitoring and biomass prediction in Colombia to support decision-making towards increased productivity, competitiveness and reduction of environmental impacts. Ten locations were sampled between 2018 and 2021 across climatically distinct areas in Colombia, comprising five farms in Patía in Cauca department, four farms in Antioquia department, and one research farm at Palmira in Valle de Cauca department. Ash content (Ash), crude protein (CP %), dry matter content (DM g/m²) and in-vitro digestibility (IVD %) were measured from Kikuyu and Brachiaria grasses during the field sampling campaigns. Multispectral bands from coincident Planetscope acquisitions along with various derived vegetation indices (VIs) were used as predictors in the model development. For each site and forage parameter, the importance of specific predictors varied, with the NIR band and Red-Green ratio generally performing best. To determine the optimum models, the effects of using a 1) averaging kernel, 2) feature selection approaches, 3) various regression algorithms and 4) meta learners (simple ensembling and stacks) were explored. Algorithms belonging to classes of commonly used models; Decision Trees, Support Vector Machines, Neural Networks, distance-based methods, and linear approaches were tested. The performance evaluation based on unseen test data revealed that CP and DM prediction performed moderately well for all three sites (R² 0.52–0.75, RMSE 1.7–2 % and R² 0.47–0.65, RMSE 182–112 g/m² respectively). The best performing models varied by site and response variable, with Regularized Random Forest, Partial Least Squares, Random Forests, Bagged Multivariate Adaptive Regression and Bayesian Regularized Neural Networks being the top performing algorithms and Random Forest Stack being the best performing meta learner. The workflow and thorough analysis of performance affecting factors presented in this study can benefit timely grassland monitoring and biomass prediction at the local level and help contribute to the sustainable management of tropical grasslands in Colombia.